JP3445394B2 - How to compare at least two image sections - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention
Words (words), multiple concatenated structures represented in an array
Components or similar units whose meaning can be interpreted.
For comparing image tokens that form
It is a law, for which characters (characters), symbols
Characters (symbols), glyphs (such as emoji), or tokens
Without individually detecting or identifying the elements that form
It's about a good way. [0002] Electronically encoded documents (electronic documents)
Text in two different formats
Often one of them. In the first format
Means that the text is a bitmap,
Text can be image data or an array of pixels
Basically defined with adjacent images of similar expressions defined
Indistinguishable. In this format, the text is a sentence
Computers that are based solely on the content of
Rarely split into image units for processing
Must. The second format is as follows
This is called the character code format.
Kist is character code (for example, ASCII code)
As a string. This character
-In code format, an image of text
Alternatively, no bitmap is required. [0003] By the optical character recognition process (OCR)
Conversion from bitmap to character code takes time
It is very expensive considering the time and labor involved in processing. Individual
Whether the character's bitmap is adjacent to it
And analyze their appearance, and furthermore,
In a group of characters preset by
Must be identified as a character. Miyatake
U.S. Pat. No. 4,956,869 to U.S. Pat.
Trace lines (equal concentration lines, contour lines)
A more efficient way to do this has been suggested. However, an original (original)
When you scan and extract an electronic document,
Depending on the image quality and noise, the actual bitmap
The appearance is uncertain. Bitmap appearance degraded
What we do is scan original documents with poor quality
Errors or affect the digital reproduction of the image
Due to similar factors. Therefore, the character
The decision-making process to identify
There is a certain uncertainty. This is of particular concern.
Means that the characters in the text are blurred,
Or it is easy to combine. Most character knowledge
Another process is one pixel connected by a certain character
Start by assuming that they are independent sets of
You. This assumption is made due to the quality of the input image.
Otherwise, character identification will fail.
You. The following patents improve character discrimination:
In particular, it demonstrates the techniques for different approaches. To Manon
U.S. Pat. No. 4,926,490 issued to Cherange
U.S. Patent No. 4,558,461 issued to Gray et al.
U.S. Pat. No. 3,295,105 to Birski
U.S. Pat. No. 4,949,392 to Long et al.
U.S. Patent No. 5,142,589 to Hood et al. The OCR method splits an image in various ways
I am trying to do it. For example, given to Chelange
US Patent No. 4,558,461 and Peppers et al.
There is U.S. Pat. No. 4,809,344 granted. [0007] In the OCR method, matching with words in a dictionary is performed.
Therefore, the reliability is improved. For example, given to Hishino
No. 4,010,445.
Fujitsu Science and Technology Journal 26,3, pages 224-2
33 (October 1990)
Is a block extraction, skew adjustment,
Split, adjacent character classification, line extraction, and
Perform pattern matching by checking and comparing with a dictionary
Show each step of character recognition by
I have. Form a string of words or characters
To identify multiple character sets,
Reading as in US Patent No. 2,905,927
Would be desirable. [0009] As a basic unit for recognition,
Using the entire code is not considered when recognizing the signature.
U.S. Pat. No. 3,1,1 issued to Frishkop
33,266. However, the split key
There is no idea to hold the character. [0010] The present invention is based on the OCR technology.
To prevent problems with
Word) and the fundamental properties of text strings
We are utilizing it. The space between words consists of letters and
Tend to be larger than the space between characters,
Separation and tokenization of tokens that make up character strings
Identify individual characters within the token
Can be improved as compared to OCR method
However, the character must be
Some correct decisions are required on the form of
Include keys such as ascenders, descenders, curves, etc.
It also includes the identification of the characters (characters),
It is easy to make mistakes. The invention, on the other hand, uses the word
Or a string of symbols or characters
The security of connected components (hereinafter referred to as tokens)
Is to be more reliably recognized and identifiable.
You. In one embodiment, the present invention provides for
Word boundaries to determine the characteristics of the list or symbol first.
Utilizes the world. Then, the toe separated within that boundary
In Kuhn's mutual or token image dictionary
A comparison with a known token is performed. So compare
Classification of tokens is not performed until the stage,
Mistakes in subsequent processing or incorrect decisions
The effects of invalid partial classification causing causation
Can be excluded. The potential of computerized text
And at least in some cases,
Processing requirements that can lead to each letter of the word
It was decided not to be imposed. So, for example
If you do a keyword search for a text image
At the time, each of the words
Instead of converting characters, one or more
Whether there is a keyword above may be defective
When continuously determining from character codes,
Instead of creating something, the pewter
Displays the shape of multiple tokens in the image and their keywords.
The token shape is compared with the token shape.
Evaluate whether the keyword exists from the ratio. This
Output to a user-friendly system
Can show some indication of the presence of a keyword with an accuracy of
Things. In addition, the new method described here
Than some methods designed to recognize lactors
It is considered that the processing speed is fast. Still further, the present invention
Can also be applied to image editing systems, and
Is not limited to the described embodiment. Characters cannot be correctly determined by the OCR method
Is likely to be relatively low, but the product
Applying the rule doubles the probability for every word
Is accumulated. Therefore, multiple words can be written using OCR.
When converted to a character code string, these
Before searching or recognizing the word
An error will occur. The present invention provides a token level
Or word level in embodiments that recognize text or text.
Pass through using divided image data
Similar to what humans use when reading and extracting text
Method enables continuous recognition. Further explanation
The process of recognizing the shape of a token
Has an effect. First, bitmap image data
Data is not lost in an irreversible condition,
Also, the reasonable display of the bitmap remains, so the user
-Characters, symbols, glyphs,
Or bitmap reproduced to determine word
You can check it. Second, use connected elements (tokens)
Each symbolic element by being
(Ie, the character (character))
Word) with its entire context and
Helps compare with token shape. For example,
Even if there are characters in the word token
Has little effect on identifying the whole word shape
And between two compared tokens representing those words
It only slightly reduces the probability of matching. Furthermore, OC
Compared to the ability of the R method, the OCR method has more characters.
It is easy to get wrong results for words you have
In contrast, the present invention generally identifies more arduous words.
Have the ability to The OCR method uses a key represented by a bitmap.
Character code, and then convert it to a bitmap
You may lose the content containing the information. Generally, this professional
Seth uses the original bitmap from the character code.
It is not a reversible thing to get a tip. However
When the word token is identified based on the shape,
As described in accordance with one of the
Can have bitmap information up to
This allows the bitmap to be reconstructed. According to the present invention, each of
Represents Tokun and has multiple image signals
Comparing at least two image sections
A method for determining similar tokens, comprising the following steps:
Can be provided. (A) An image representing the first token
Storing the signal in a first model memory;
The first token is expanded and represented in the image memory
And (c) an image sig representing the second token
(D) storing the null in the second model memory;
Inflated representation of the second token in the image memory
And (e) storing the image stored in the first model memory.
The image signal stored in the second image memory
Image signals to determine the first similarity distance.
(F) the image stored in the second model memory
The image signal stored in the first image memory.
Compare image signals to determine second similarity distance
And (g) the distance between the first and second similarities
In response, determine whether the first and second tokens are similar
judge. One aspect of the invention is that each image section is
Represents a coon, a few with multiple image signals
Compare at least two image sections to see similar
A method for identifying a token, comprising: (a) a first token;
The image signal representing the image in the first model memory
(B) storing the first talk in the first image memory;
(C) The second talk
The image signal representing the
(D) storing the second toe in a second image memory;
Creating an expanded representation of the kun, and (e) the first
The image signal stored in the model memory,
The image sig stored in the second image memory;
Comparing with null to determine a first similar distance;
The image signal stored in a second model memory;
The image stored in the first image memory.
A second similar distance is determined by comparing the
(G) corresponding to the first and second similar distances
Whether the first token is similar to the second token
The ratio of at least two image sections
It is a comparison method. BRIEF DESCRIPTION OF THE DRAWINGS FIG.
What is shown is to illustrate a preferred embodiment of the present invention.
And not to limit it to equivalents. FIG.
Outlines a generalized image processing system.
This covers many situations where the invention can be used effectively.
-I can do it. Generally, the source image is
Sources such as facsimile machines or recording systems
It is extracted from the image extraction system 2. This source image
The page is sent to the computer processing unit 4, but the processing unit
4 can be any of several known devices, described herein.
Such a device according to the invention may be used. User interface
In response to the command input to the source 6, the processing device 4
Outputs to an output device 8, which is also a printer,
Display, facsimile machine or other recording device
It may be. Basically, as shown at the top of FIG.
The input document is put into the system and the output document is
Is collected. In the following, an image is an image bit.
Is described as a map, and here the image
What are multiple rasterized (rasterized) (scan lines
(Resolved into the image signal). This
These signals are usually called pixels and correspond to the text.
When expressing a mark or active position
Commonly appear in black, which creates documents and marks
Is done. These configurations are used to describe the present invention.
But limited to black and white or binary images
It is not done. Rather, the present invention represents an image
It is generally applicable over a wide range of technologies.
In addition, the present invention provides for intra-image or inter-image
It also aims to determine the similarity of multiple tokens
You. In one embodiment, the present invention provides a simple
Determine similarity of word objects (word objects)
This is useful for editing and compressing images.
It can also be used for
It is not completely limited to the examples. FIG. 2 shows a word which is an embodiment of the present invention.
A system for determining, dividing, and comparing shapes is shown.
You. Each element of the system is a number of devices
Or just a single device in one device.
It may be a program. Similarly, in the following the word
Describe the preferred embodiment for recognizing objects
However, the comparative technology that forms the basis of the present invention
Requires strict pre-processing operations as described in the examples
It does not do. Triggered by input bitmap 10
But this source is definitive or part of the present invention
It does not make up. Bitmaps are initially partitioned
(Segmentation system) 12
With multiple tokens (words, character strings,
Or other units whose meaning can be interpreted)
Is done. First, the image bitmap is
-(Skew return unit) 14
Determines the angle of text oriented in the image, and
Correct the orientation of. Made by this deskew operation
Word boxer using the deskewed image
(Word enclosure) at 16
The boundaries are determined, along with the boundaries of this token, the image
The boundaries of the text lines within are also identified. Word separator
(Word segmenter) 18
Word token boundaries are applied to the map and
Each word group in the image is read
Are separated in order and these are then
Will be handled. Here "word (word)", "symbol
String "or" character string (character
Column) "means connected alphabets or punctuation
Element, or a broader set of tokens
And form all or some of the units whose meaning can be interpreted.
Is what it does. Such interpretable units are:
Characterized in the image, the unit itself
Of adjacent elements, signs or symbols
The gap is larger than the gap to be separated (spacing).
Have been separated. In this regard, the invention can be applied differently.
Yes, for example, in text and word editing systems,
Independent words to process (operate) the image continuously
Can be used. Therefore, the present invention is related to word recognition.
However, the present invention is not limited to this. Next, the shape comparator 24 determines whether the
The shape of the word token representing each word is stored in the dictionary 26
Form of a known or already identified word token from
Compare with the shape. As another example, the shape comparator 24 is
Two or more words determined from page 10
It may be used to compare the shapes of the bubbles. desirable
In an embodiment, the comparator 24 is
To characterize the similarity between
Using the Hausdorff distance as a variable
You. More importantly, shape comparator 24 is identified
Word Talk from Strings of Non-Characters
Only compare the token shape to a known word token shape
It is inconvenient. In a simple context,
The converter 24 converts the shape of one token
It is a simple device to compare with the shape,
Is represented by the matching instruction output. Two
The similarity between token shapes is relatively indicated. The token shape in the context of the word image is
To outline the method or device to be determined and compared
Next, each step of the embodiment for comparing the shapes will be further described below.
explain in detail. To further explain the process of the present invention
FIG. 3 shows a sample image, which is a public image.
Are taken from what became the assets of
Several lines are included. Figure 3 shows a text (text) page.
The outline of the appearance of the image on the page is shown in FIGS.
And 6 are part of the scanned image of the page
Where the bitmap image is enlarged
To illustrate the problems with known OCR technology.
You. Referring to FIG. 3, for example, 2 of this text image
The word image 50 on the line is “formation”, and 4
The word image 54 on the line is “automobile”
Some characters appear to be connected. In addition, many known small angle images
How to rotate or correct skew
Used to get a deskewed display of the image
be able to. When the image is deskewed, the token
Can be extracted in many ways, and the choice of
Depends on the first application (application) of Kung comparison
You. The comparison technique illustrated in the present invention is based on multiple
Used for multiple component tokens represented within boundary 58
That they match each other, or that
Identify a match with Kung. Once identified, the larger
Matched or known tokens in the sentence image
Labeled or similarly for subsequent processing
Be identified. For example, for subsequent processing,
Identify, access, and extract information in protected documents
And publicly announced on June 24, 1993
Peter B. opened. Published by Mark et al. (W0-93)
/ 12610) "A method for compressing images and
The compression technique disclosed in "Apparatus" is also included. Docume
Multiple tokens representing parts of the
Can be used, but with advanced image processing, word-based
Generating a kun is desirable for identifying words.
Structured by words and strings of related characters
An example of comparing and recognizing multiple tokens generated is as follows
Will be described. As shown in FIG. 2, the word boxer 16
Are shown in FIGS. 7 and 8 for the deskewed image.
An operation is performed according to the flowchart shown. Word boxer
The following description of the process steps performed in the
Using operations performed on a functional computer
However, since the present invention is limited to this embodiment,
There is no. When step 80 starts, the word boxer
First, read the input image in Fig. 3 and this image is necessary
If there is, it is deskewed by the deskewer 14. This
Functions, for example, a hard disk or similar storage
Image stored in memory such as storage device
Access to the image, and furthermore, the image
Page of memory allocated for that image.
Copy it to a location and, if necessary, its image
Assign a pointer to. When the image is extracted (retrieved), the step
Step 82 finds the connected element in the image
You. Is this process inside a stored binary image?
Only find the black pixels. Black pixels are found
The interactive (interaction) process continues to
Black pixel and the adjacent black pixel
Find them and continue until the range of connected pixels is determined.
Do it. More specifically, the eight-neighbor definition (eight-n
eighbor connection definition) is used. Sand
That is, one pixel has eight compasses (ranges)
If they are adjacent to one of the directions, they are adjacent
It is considered to be the same connecting element. Furthermore, the image
All black pixels within are properly associated with other black pixels
Until the connected elements are formed
Repeated. As shown in FIG. 4, the connected pixels
When attached, a square box or border 58 is identified
Which reflects the largest range of connected pixels
And the square box represents the xy coordinates of the image.
Oriented along. The glue of all connected elements in the image
Box around bounding box (bounding box
Is defined, a part of the image is shown in FIG.
Boxes or boundaries of connected elements, as shown
Out of the set where the word boxer is bad
Find the box (not shown) (not shown)
put out. Bad box is characterized as follows
You. (A) A tall box whose height is
About 20 percent higher than the height of the image
More than 90 percentiles in the page
Big box or (b) short box
And less than almost 1/3 of the height of 90 percentile
Box. After the analysis, the remaining boxes are
The vertical or y axis of the document (the y axis is the deskew
Considered to be the axis perpendicular to the direction of the line of text
Projected to form a histogram, which is a box
Number of borders of a reflection as a function of position along the y-axis
As shown in FIG. 9 for all the images in FIG.
You. In the preferred embodiment, the hiss projected on the y-axis
Determine the position of the text line for the program data
Before smoothing (smoothing)
May be performed. Next, from the result of the histogram,
Provisional lines or line boundaries are aligned along the y-axis of the image.
This line is
It is a valley in Lamb. For example, as shown in FIG.
In addition, a plurality of valleys or lowest points 120 are adjacent peaks
Alternatively, it is identified between the highest points 122,
The position of the space between the lines can be determined from 120,
These are shown as reference numeral 62 in FIG. This operation
This is performed by step 88. Finally, provisional (preliminary)
Once a typical text line or row is determined,
Assign all connected element boxes to the specified line
Function operates. If the position of the text line or line 62 is temporarily
Once decided, the bounding box of the connected element
First, the two rows that lie on a particular row
Is performed. Flow chart steps
Steps 92, 94, and 96
In the meantime, the temporary
Check if the regular text line is correct
Will be First, as explained later, text
Function to confirm that line separation is not a failure
You. In general, focusing on connected elements, a certain text
The projected parts of the rows are as bad as they are in the y-axis direction.
Unless they overlap, the part projected in the x direction
They do not overlap. Identified in step 92
If the projected parts overlap,
Rows may be two or more separate rows
Look for the lowest point in the higher, y-projected graph.
Must be attached and separated. Also, text image
For example, a dot above "i" or an underline of a word
Around a small group of connected elements
Ignore text boxes and separate text lines further
Must not trigger the trigger by mistake
Must. Second, as in step 96, x
The remaining boxes that overlap each other along the axial direction
One button with a border surrounding the merged element
Merge into the box. Generally, this merging process
Overlook multiple boxes in one line and overlap in the x direction
And in the y direction
Identify the box. The minimum overlap in the y direction is about 5
About 0% is good. For example, a scanned image
If the message contains the word "fort", scan
Will cause the right edge of the "f" box to
May overlap the left edge, and therefore overlap along the x-axis.
Merged box elements will result in the “f” and “o”
Will be merged. In this procedure,
A size test is also performed, and a box smaller than the specified size is
Boxes are not merged. Then this little box
Can be identified and removed as noise in the image.
Wear. Third, text lines are correctly detected
And the remaining boxes in this row are connected elements or
Tokens, some of which are words or
To form similar elements that can interpret meaning
Further connection is required. Combine further adjacent elements
-Based talk in scanned images
In step 98, a text line is
Histogram of separation distance between adjacent elements
Is performed continuously. Diagram showing the distribution of common text lines
10 and the dashed curve is the histogram data of the row.
And the solid curve is a smoother version
is there. As expected, the resulting curve shows the distribution of the binary model
In general, the first set of peaks 130 and 132 is
Represents the distribution of spacing separation between characters
In contrast, the second peak is broader and more frequent.
Low, reflecting the separation between adjacent words
I have. In addition, under certain conditions, the distribution of a single model
It is. Two maximum values of the distribution of the binary model (bi-model)
First sets the separation threshold in step 100
Used to identify, followed by separation between words
Separation and separation between characters
Also used to distinguish. Next, using this threshold value for separation,
Separation in the x direction among adjacent boxes in the text line
Merge smaller than the separation threshold
To do so, the procedure of step 102 of FIG.
You. In this step, simply concatenate each row
Set of elements that are adjacent to each other
All that are separated by a distance shorter than
Is done. Merge adjacent characters in a word
And the resulting box structure contains each text
Reflects the boundaries of word tokens within a line, for example,
FIG. 6 shows a plurality of boxes 66 surrounding a plurality of words.
It is. At this point, a small unmerged Bock
Option to recognize and remove noise as noise in the image.
May be performed. Next, read in order (from the top
Bottom, and from left to right of each line of text)
A list of the created boxes is created at step 104.
Each of the arrays in the box list is
One word token, picture, phrase in each
Readings or bows of equivalent units that can interpret meaning
A binding box 66 is defined. Returning to FIG. 2, for example,
A box list that represents the boundaries of the domain-based tokens
The spelling of a token by the word boxer 16
Once the list is created, this list and bitmap image
Page is a token or word segmenter
18 is sent. Generally, a divider (segmenter) 18
Is an image processing device, and the input image
The bitmap of page 10 is specified in the box list.
A series of smaller words according to word or token boundaries
It can be divided into various bitmap images. Wah
The output from the dosegmenter 18 is a bitmap image
Is a series of flows (series), each image
Is the word toe identified by word boxer 16.
Bean or equivalent
It has a bitmap. In a preferred embodiment,
The word segmenter 18 is surrounded by a word box.
For each part of the input image
It does not actually generate a bitmap. Rather,
The segmenter can simply open the window or
By selecting certain parts of the
The image specified to be within the bounds of the token box
Access to that part of the page. Destination
As described above, the output of the word segmenter 18 is
Is sent to the comparator 24, where the
And the token is another bitmap image from the dictionary 26
And the token output from segmenter 18
The image matches the word token supplied from the dictionary
It is determined whether or not to do so. As described below, the word image
Housed as one preferred way to compare professionals
A technique for measuring Loff distance is used, which is 1991
"Housedorf
Comparison of Images Using Distance "(TR91-121
1), and the December 1992 "Housedorf
Multiple Resolution Technology for Comparing Used Images "(TR
92-1321), all of which are
Published by Computer Science Faculty at Nell University
ing. Generally, boxed tokens are
The method of comparison is to identify the element identified in a particular box.
11 and 12 for comparing elements.
Process is used. Simplified implementation described below
An example is the same word-based token in an image
Or different. It
The bitmap of each section or word token
Expression by a loop is used to define predetermined boundaries.
Defined by a box (bounding box)
It corresponds to the area. Between bitmap sections
A common way of making such comparisons is
Generally known, the logical AND of two images
(Logical AND) operations are used to determine similarity.
It is. In the present invention, on the other hand, this correlation is expanded (di)
lation) technology to improve
Specific to the digitization process used to form the image
The effect of the quantization error is eliminated. As adopted below, the two compared
Two token images as Box 1 and Box 2
I do. These image parts (image sections)
Is even if two sections from the same image
Good, even two sections from different images
Or one section from an image,
The string, word or toe of the entered symbol
Electronically from an interpretable unit of meaning forming a cun
One created section may be used. Figure 2
It is represented as a word image "dictionary",
The general purpose of block 26 is to use other token images
Token to compare with section (box 1)
By providing a section of the image (Box 2)
is there. As shown in FIG. 6, "automobile" 70 and
Original word representation of two word-based tokens, 72 and 72
May be compared according to the description, where the expression 72 is
Book. Box 1 and Box 2
Once sections 70 and 72 are defined, respectively,
Each image is called a "model".
An inflated version of this model is made,
Image ". As shown schematically in FIGS.
The comparison method used in the comparator 24 is as follows.
Pixel in model 1 (150), ie, box
Replace the original pixels in the section surrounded by
Image 2 of the expanded representation of the pixel represented by box 2
The pixel is compared with the pixel of (156).
A first distance is generated. Similarly, this process
Box that is inverted and is a pixel in model 2 (152)
The original pixel of the section surrounded by 2 and the box
Image 1 (15)
4) is compared with the pixels in
Generated at block 162. Subsequently, block 16
At 4, 166 and 168 these two distances are numerical
Processed in box 1 and box 2
Determine the degree of similarity between these two image sections
Is done. More specifically, the comparator 2
4 first places image 2 in the memory location for the model
A certain word image specified in the “dictionary” (Box 6)
2) Duplicate the pixels within the boundary. These pixels are
Is model 2. The comparator then goes to memory
Duplicate Model 2 at location 2 and add
Image 2 (expanded image 156)
Is created in step 200. That is, the model 2
For all "on" or black pixels stored in Molly
On the other hand, in the memory of the image 2, the part adjacent to this periphery
Turn minute on or black. Exact numbers next to each other are half-expanded
Defined as the diameter (delayion radius)
Is predetermined. As an example, the preferred expansion radius
Is 1.0 pixels, four adjacent pixels are turned on, and
Assuming that the diameter is 1.4 pixels, all eight neighboring pixels are adjacent.
Turn on. If the expansion radius is further increased, the same
Non-words may be incorrectly matched
growing. The choice of the expansion radius prevents errors by quantization.
This error is mainly due to digitization.
Occurs in the process. When choosing the expansion radius,
When calculating a simple correlation (effectively increasing the expansion radius to 0
Errors that would be introduced in
Is desirable, while images that are over-inflated (eg,
(For example, a large expansion radius)
Rebellion must be avoided. Therefore, pixels 1.0 and 1..
A desired expansion radius of four ranges is acceptable within this limit
Shown as a compromise. Next, the process determines the symbols to be compared.
Model and inflated image versions for strings
Repeated to make. For example, the whole input image
Ten copies may be inflated as described above,
In the box 202, the information specified in the box list
Pixels within the expanded border for all boxes are
Copied from the dilated input image. These pixels
Represents an individual dilated "word", and
At image 1 (the dilated portion of the input image,
156), whereas the original of the input image
Word segments not expanded by
Dell 1 (150). Same as box 2 image
The pixels representing the words in each image are fat
It looks and is even more compact than the corresponding model. Input image and image of "dictionary"
Image and their associated models and inflated images
Created and stored in memory, a pair of inputs (box
1) and dictionary (box 2) images are for comparison
Is selected in step 204. Next, Compare
Data 24 in step 206
Test to determine if the
Do That is, these boxes are
Whether the height and height are within a predetermined range.
You. As shown in FIG. 11, the dimensional difference ΔL
ΔL = | L1-L2 |
You. Further, in the size test in step 206,
It is desirable to do a height comparison (not shown)
This comparison is done in the same way as described for the length comparison.
be able to. Within these boxes, each other
By contrast, the image may be shifted,
Large dimensions to further improve reliability
A difference in the law may be allowed. Dimensional difference ΔL is within specified range
If not, at step 208 more buttons
(Image section) is available
Then, in step 204, a different image
Pairs (input and dictionary) are selected for comparison.
Otherwise, the image of the selected input and dictionary
Boxes indicating the boundaries of the diaper are approximately the same size
Each word box pair
Are further compared to see if they match
You. A binary image is a limited set of points A
It is thought that it represents the point, the position of each point of A
The target is represented by the "on" pixel in the binary image.
It is. Therefore, the measurement for the set of points to be compared
Degrees, Hausdoff distances to compare binary images
It is considered applicable. In particular, the limitations of A and B
Given a set of points
It is defined as follows: [Outside 1] And | ab | are two given points a and
b. In practice, the function h (A, B) is
The distance of A to the nearest point of B
Source, and the largest ranked source
The distance value at these points
It is specified. Therefore, for h (A, B) <= δ (delta)
If there is, each point of A is a point of B
Means within the distance δ. The function H (A, B) is
Shows the maximum of the two asymmetric distances, and therefore H (A,
B) If <= δ, each point of A is
Point within δ, and vice versa. This c
The Usdorf distance is thus two binary images (there are
Or a limited set of points)
Degree), and if the value of δ is large,
Indicate low similarity. Compare bitmap images of tokens
When comparing, the small value of δ
Quantization noise (random at token boundaries
Or off pixels).
However, these images may be relatively similar.
Can be. calculate the Hausdoff distance for small values of δ
As a desirable method, an operation to take a logical AND (logical product)
The expansion technique (delayion) is used together with the work. Ah
When the binary image A is expanded by a certain radius δ,
Each on or black pixel of image A has a radius δ
Has been replaced by a circle. The four most of a pixel
To represent neighbors (horizontal and vertical), δ = 1.
0 is used, while δ = 1.4 is equivalent to the 8
Represents the nearest (horizontal, vertical and diagonal). This
These are desirable values to cancel the quantization noise.
You. B ′ is obtained by expanding A by δ.
And h (A, B) <= δ is A 正 し く B '= A correctly
It is clear that the mushroom is where ∧ is the theory of A and B '
Represents logical product (AND). That is, all the black spots of A are B
Must be within the distance δ of any sunspot of
All black spots in A must match certain black spots in B '
Absent. Therefore, it is determined whether h (A, B) <= δ
Similarly, H (A, B) <= δ simply replaces B with δ
Only inflated (estimated from A) and further A (estimated from B
) Is calculated. In general, any point of A is
May not be close to any point, and vice versa
You. Therefore, it is necessary to find the maximum value of the Hausdoff distance.
Amount (eg, median or other percentile)
It is common to replace it with calculations. This definition is
is there. [Outside 2] This calculation calculates the nearest point of B from each point of A
No. K instead of the maximum (longest) distance to the point
The eyes are calculating large values. So there is some number of A
Or some parts are ignored in the calculation of this distance.
If there are m points in set A, let k = m
Then, this definition is the same as h (A, B). But one
Generally, for a certain value of τ that takes a value in the range of 0 <= τ <= 1,
Assuming that k = τ × m, mk = (1-
τ) × m is ignored (ie, at points near B)
It doesn't have to be). This also applies to H (A, B)
it can. In the preferred embodiment, τ is 4 percent of A
And may not coincide with B ′ (τ> = 0.
96), and vice versa. This calculation is preferably applied to small values of δ.
In the case of
There may be a point of A that does not overlap with B '(and vice versa)
In Although). These non-overlapping points are defined.
Not less than fraction τ
must not. Therefore, the black images in A and A∧B '
We calculate h '(A, B) by comparing prime numbers.
You. Let p be the number of black pixels in A, and q be A∧B '
Given the number of black pixels, for a given value of τ, exactly
When h ′ (A, B) <= δ, q / p> = τ. Further, the present invention provides a method for calculating the Hausdorff distance.
A and to find the small value (best alignment)
Hausdorff when B shifts relative to each other
It is also possible to evaluate the distance. This technique is known in the art.
Relative shifted used in relational operations
Same as the case, except that this reduces quantization noise.
Finding correlations that are important in terms of minutes
And clearly different. To find the correlation,
The point between point B and point B
(For example, using δ = 0,
Case is closer to a correlation). To execute the above-described technique,
Using a generalized procedure, two image sections
Can be compared to determine if they match. 1) Model 1 is superimposed on Image 2
You. 2) Black model 1 image that matches the pixels of black image 2
Count the number of primes, then divide by the total number of black model 1 pixels
(Step 214) 3) The percentage of matching black pixels is a predetermined threshold
Percentage of value τ (τ is preferably about 0.96)
If above, these boxes are the first verification
In the example), it is determined that they match (step 216). 4) The model 2 is superimposed on the image 1. 5) Add these two image sections to the above steps
The second pixel of the matched black pixel is compared again as in
Determine the percentage (step 220) 6) This second percentage is the percentage of the predetermined threshold
Above the percentage τ, these boxes are
Is determined to be consistent with the verification (example)
222). And 7) these image sections in both tests
If they match, consider them to be the same word
In addition, the words from comparator 24 in FIG.
The provided instruction is output (step 224). In addition to the embodiments described above, the present invention
Word talk in an image that can be used as a comparison technique
You can create an equivalent class for The invention uses OCR
It can be used as a pre-processing operation of the system,
This allows for the speed and precision of the available OCR system.
The degree can be improved. As yet another alternative,
Akira decides that word tokens appear repeatedly
Can also be used to
Is converted to a certain icon to reduce the size and large text
The overall size of the data file needed to hold it
Can be reduced. In this preferred embodiment, the box
Compare the sizes and boxes or images that might match
Existing classes in the page section (for example, input image
And the same token in both
Create a library for each of several boxes
Program is executed. For example, their length
Classify token image section by (width)
Data structure suitable for
Can improve the speed at which pairs are compared. A certain input
Some parts of the image are replaced with known or "dictionary" tokens.
Although comparison has been described, for example, FIG.
In image sections 70 and 72, the present invention is the same.
Compare tokens in one or different images
Are possible and have been described for the purpose of illustrating the operation of the present invention.
It is not to be understood that it is limited to examples. According to the present invention, a plurality of image signals or
Represents two image sections composed of pixels.
Or a method of comparing tokens, where each token
Coons are represented by one or more linked symbols.
And are identified as the same token. The present invention
Is also the symbol or character that forms the token.
Operate without the need to individually detect or identify
You. This method detects elements in an image and
The token boundaries, then apply a two-step process.
Model in which the inflated image represents the token
And the relative similarity between them is determined.
It is. Thus, according to the present invention, elements or similar
Multiple toes that define an image area with
Obviously, we can provide a way to compare kung. This
The invention is described with reference to the preferred embodiments.
To make it available for use in computer systems.
Described as measured software means,
One or more microprocessors capable of executing commands
Use a processor or a processor with computational capabilities to
Explained above about image data processing
Such operations can be performed. In addition,
The description is based on features designed to perform the processing described here.
It can also be realized using fixed hardware. In addition,
Ming explained as part of a larger word recognition system
is there. However, as noted above, the present invention provides
Or image editing or related systems
It can also be used for programs. In fact, the token
Or identify strings of symbols, classify them,
Book on any system that needs to be grouped.
The invention can be used. Finally, the present invention relates to text
The explanation is based on the format image. But the text
For images that contain non-format images
The same can be applied.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a schematic system diagram of an image processing system that can effectively use the present invention. FIG. 2 is a blockgram showing combinations of system elements that make up an embodiment of a novel token recognition system applicable to word images. FIG. 3 is a diagram showing an image sample from which text is extracted as an example to explain a new process. FIG. 4 shows a portion of a scanned image of an example text at an intermediate stage in the process of the present invention. FIG. 5 illustrates a portion of an example scanned text image at an intermediate stage in the process of the present invention. FIG. 6 illustrates a portion of a scanned image of an example text at an intermediate stage in the process of the present invention. FIG. 7 is a flow chart illustrating a process for determining word boundaries in an image. FIG. 8 is a flow chart illustrating a process for demarcating words in an image. FIG. 9 is a graph showing histogram data obtained in step 87 of FIG. 7; FIG. 10 is a graph showing histogram data obtained in step 98 of FIG. FIG. 11 is a flow chart with illustrations illustrating processing operated by the comparator of FIG. 2 in accordance with the present invention. FIG. 12 is a flowchart outlining a process for comparing images within word boundaries shown in FIG. 11; [Description of Signs] 2 Source 4 Image processing 6 User interface 8 Output destination 10 Input image 14 Deskewer 16 Word boxer 18 Word segmenter 24 Word comparison "Dictionary" of word image

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-5-242298 (JP, A) JP-A-7-200732 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G06K 9/00-9/82

Claims (1)

(57) A method for identifying at least two image sections, each image section representing a token, and comparing at least two image sections with a plurality of image signals, to identify similar tokens, (A) storing an image signal representing a first token in a first model memory; (b) creating an expanded representation of the first token in a first image memory; Storing an image signal representing the second token in a second model memory; (d) creating an expanded representation of the second token in a second image memory; and (e) generating the first model. Comparing the image signal stored in a memory with the image signal stored in a second image memory to determine a first similar distance; Comparing the image signal stored in a model memory with the image signal stored in a first image memory to determine a second similarity distance; and (g) the first and second similarities. Determining whether or not the first token and the second token are similar according to the distance of at least two image sections.